FastWan 2.2 5B — pinned Studio model bundle
Public mirror of FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers at 3e187042a324f6f5fb68fd22110a78725253de8f.
The model weights, tokenizer and configurations are unchanged. See
UPSTREAM_README.md for author attribution and
provenance.json for original file hashes and sizes.
Runtime
FastWan uses three-step DMD sampling (timesteps 1000, 757, 522, guidance 1). It requires the FastVideo DMD scheduler or an equivalent implementation; an ordinary Wan scheduler is not a drop-in substitute. CPU offload and VAE tiling are recommended on smaller GPUs. Model weights alone are not a Runpod worker. This repository contains no deployment credentials or private app source.
Resolution and duration
The upstream training profile is 1280 x 704, 121 frames at 24 fps. Alternative resolutions may reduce quality. Studio's operational range is 128–1280 pixels per side in multiples of 32, at most 901120 pixels, and 5–121 frames in the form 4n+1. This range is an application limit, not a claimed model maximum. Only 512 x 288, 25-frame text-to-video and image-to-video smoke outputs were verified. Their motion and detail were rough; larger settings are unbenchmarked.
Adapters
Speed distillation is already baked in. Do not stack the extracted FastWan speed LoRA on top. No additional realism or artistic-nudity adapter has been verified or bundled. Wan 14B adapters are not compatible with this 5B model. This is a faithful backup of the selected fast model, not a claim that it outperforms Wan 14B in realism, motion, or speed.